Persona-German / README.md
Ill-Ness's picture
Update README.md
afdcfb2 verified
metadata
license: cc-by-4.0
task_categories:
  - text-generation
language:
  - de
tags:
  - synthetic
  - personas
  - german
  - custom
  - datadesigner-inspired
size_categories:
  - n<1K
dataset_info:
  features:
    - name: uuid
      dtype: string
    - name: professional_persona
      dtype: string
    - name: sports_persona
      dtype: string
    - name: arts_persona
      dtype: string
    - name: travel_persona
      dtype: string
    - name: culinary_persona
      dtype: string
    - name: persona
      dtype: string
    - name: cultural_background
      dtype: string
    - name: skills_and_expertise
      dtype: string
    - name: skills_and_expertise_list
      sequence: string
    - name: hobbies_and_interests
      dtype: string
    - name: hobbies_and_interests_list
      sequence: string
    - name: career_goals_and_ambitions
      dtype: string
    - name: sex
      dtype: string
    - name: age
      dtype: int64
    - name: marital_status
      dtype: string
    - name: education_level
      dtype: string
    - name: bachelors_field
      dtype: string
    - name: occupation
      dtype: string
    - name: city
      dtype: string
    - name: state
      dtype: string
    - name: zipcode
      dtype: string
    - name: country
      dtype: string
    - name: language
      dtype: string
    - name: persona_region
      dtype: string
    - name: profile_tag
      dtype: string
  splits:
    - name: train
      num_bytes: 70746
      num_examples: 240
  download_size: 70746
  dataset_size: 70746
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Persona-German

Ein deutsches, synthetisches Persona-Dataset im Stil von Nemotron-Personas-Korea
A German synthetic persona dataset inspired by Nemotron-Personas-Korea

Dataset Overview

Persona-German is a custom synthetic German-language persona dataset built in the style of nvidia/Nemotron-Personas-Korea, but adapted for a smaller local dataset release. It contains German persona descriptions together with demographic and geographic context fields so the data can be filtered, conditioned, or expanded for downstream training and evaluation workflows.

The current release contains 240 records stored in 2 parquet files. Each record includes multiple persona views such as professional, sports, arts, travel, culinary, and a concise overall persona summary. Additional attributes such as cultural background, skills, hobbies, goals, age, education, occupation, and location are included to make the personas easier to target in synthetic-data or LLM experiments.

All records in this dataset are synthetic. The generator was written locally and includes a sentence-level deduplication check so the same sentence is not repeated across the persona text fields more than once.

--

Persona-German ist ein benutzerdefiniertes synthetisches deutschsprachiges Persona-Dataset, das sich strukturell an nvidia/Nemotron-Personas-Korea orientiert, aber als kleinere lokale Veröffentlichung für deutsche Personas umgesetzt wurde. Es enthält deutsche Personabeschreibungen zusammen mit demografischen und geografischen Kontextfeldern, damit die Daten für Training, Evaluation oder konditionierte Generierung gefiltert und erweitert werden können.

Die aktuelle Version enthält 240 Datensätze in 2 Parquet-Dateien. Jeder Datensatz umfasst mehrere Persona-Perspektiven wie Beruf, Sport, Kunst, Reisen, Kulinarik sowie eine kurze zusammenfassende Persona. Ergänzende Attribute wie kultureller Hintergrund, Fähigkeiten, Interessen, Ziele, Alter, Bildung, Beruf und Wohnort erleichtern die gezielte Nutzung in synthetischen Datenpipelines und LLM-Workflows.

Alle Einträge in diesem Dataset sind künstlich erzeugt. Der lokale Generator enthält außerdem eine Duplikatsprüfung auf Satzebene, sodass derselbe Satz in den Persona-Textfeldern nicht mehrfach vorkommt.

What Is Not In The Dataset

This dataset is not an official NVIDIA release and is not a replication of the real Nemotron-Personas-Korea production pipeline. It does not include family, housing, district-level, or military-status fields because those columns are not present in the generated parquet schema for this local version.

The dataset is also much smaller than the Korean reference dataset and should be treated as a custom seed dataset or scaffold rather than a population-scale release.

--

Dieses Dataset ist keine offizielle NVIDIA-Veröffentlichung und bildet die Produktionspipeline von Nemotron-Personas-Korea nicht vollständig nach. Familien-, Wohnungs-, Militärstatus- oder kreisgenaue Regionalfelder sind hier nicht enthalten, weil diese Spalten im tatsächlich erzeugten Parquet-Schema dieser lokalen Version nicht vorkommen.

Außerdem ist das Dataset deutlich kleiner als das koreanische Referenz-Dataset und sollte eher als benutzerdefinierter Startpunkt oder als lokales Persona-Scaffold verstanden werden, nicht als bevölkerungsnahe Großveröffentlichung.

Data Developer

Custom local dataset created in this workspace.

Release Date

Released locally on April 26, 2026.

Dataset Creation Date

April 26, 2026

License / Terms of Use

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

--

Dieses Dataset steht unter der Creative Commons Attribution 4.0 International License (CC BY 4.0).

Use Case

This dataset can be used for:

  • German persona generation
  • Synthetic data bootstrapping
  • LLM fine-tuning experiments
  • Prompt conditioning and retrieval tests
  • Evaluation of persona diversity in German outputs

--

Dieses Dataset kann unter anderem für folgende Zwecke genutzt werden:

  • deutsche Persona-Generierung
  • Bootstrapping synthetischer Daten
  • Fine-Tuning-Experimente für LLMs
  • Prompt-Konditionierung und Retrieval-Tests
  • Evaluation der Persona-Vielfalt in deutschen Modellausgaben

Data Version

1.0 (04/26/2026)

Intended Use

Persona-German was created as a compact German counterpart inspired by the structure of Nemotron-Personas-Korea. It is intended for experimentation, prototyping, and small-scale dataset development where German-language personas with consistent structure are useful.

Because the dataset is synthetic and relatively small, it should not be treated as a faithful statistical representation of the entire German population. It is better suited to controlled testing, synthetic prompt generation, schema prototyping, and downstream expansion than to demographic research.

The dataset includes only adult personas aged 19 and above.

--

Persona-German wurde als kompakte deutsche Variante im Stil von Nemotron-Personas-Korea erstellt. Es ist für Experimente, Prototyping und kleinere Datensatz-Workflows gedacht, in denen strukturierte deutschsprachige Personas hilfreich sind.

Da das Dataset synthetisch und relativ klein ist, sollte es nicht als exakte statistische Abbildung der gesamten deutschen Bevölkerung verstanden werden. Es eignet sich eher für kontrollierte Tests, synthetische Prompt-Erzeugung, Schema-Prototyping und spätere Erweiterungen als für demografische Forschung.

Das Dataset enthält ausschließlich erwachsene Personas ab 19 Jahren.

Dataset Details

The dataset contains:

  • 240 records across 2 parquet files
  • 26 fields: 6 core persona fields, 4 additional persona attribute text fields, 15 demographic and contextual fields, and 1 unique identifier-style tag
  • German-language persona descriptions
  • Adult personas only
  • Sentence-level deduplication across persona text fields

--

Das Dataset enthält:

  • 240 Datensätze in 2 Parquet-Dateien
  • 26 Felder: 6 zentrale Persona-Felder, 4 ergänzende textuelle Persona-Attributfelder, 15 demografische und kontextuelle Felder sowie 1 identifier-ähnliches Tag
  • deutschsprachige Personabeschreibungen
  • ausschließlich erwachsene Personas
  • Satzebenen-Deduplizierung über die Persona-Textfelder

Schema

The dataset includes the following columns:

  • uuid
  • professional_persona
  • sports_persona
  • arts_persona
  • travel_persona
  • culinary_persona
  • persona
  • cultural_background
  • skills_and_expertise
  • skills_and_expertise_list
  • hobbies_and_interests
  • hobbies_and_interests_list
  • career_goals_and_ambitions
  • sex
  • age
  • marital_status
  • education_level
  • bachelors_field
  • occupation
  • city
  • state
  • zipcode
  • country
  • language
  • persona_region
  • profile_tag

Seed And Generation Notes

The dataset was generated from local templates and controlled attribute pools in generate_persona_german.py. It does not claim alignment to official German census distributions. Instead, it follows a structured synthetic generation approach with rotating attribute pools and schema-consistent output suitable for local experimentation.

The generator also validates that no full sentence is repeated across the main persona text fields.

--

Das Dataset wurde aus lokalen Vorlagen und kontrollierten Attribut-Pools in generate_persona_german.py erzeugt. Es erhebt keinen Anspruch auf eine exakte Ausrichtung an offiziellen deutschen Bevölkerungsverteilungen. Stattdessen folgt es einem strukturierten synthetischen Generierungsansatz mit rotierenden Attribut-Pools und schema-konsistentem Output für lokale Experimente.

Der Generator validiert außerdem, dass in den zentralen Persona-Textfeldern kein vollständiger Satz doppelt vorkommt.

How To Use It

from datasets import load_dataset

persona_german = load_dataset("8F-ai/Persona-German")

If you are loading it locally from files instead of from the Hub:

from datasets import load_dataset

persona_german = load_dataset(
    "parquet",
    data_files={"train": "Persona-German/data/train-*.parquet"},
)

Dataset Characterization

Data Collection Method

  • Synthetic

Labeling Method

  • Not applicable

Dataset Format

  • Text
  • Parquet

Dataset Quantification

  • Number of records: 240
  • Number of parquet files: 2
  • Total storage size: 70,746 bytes

Ethical Considerations

This dataset is fully synthetic, but synthetic personas can still encode design biases from template choices, category pools, and coverage decisions. Users should review the data before using it in production systems or in any application that may affect real people.

The dataset should not be interpreted as a statistically complete or authoritative representation of Germany, German regions, or any demographic group.

--

Dieses Dataset ist vollständig synthetisch, dennoch können auch synthetische Personas Verzerrungen aus Vorlagen, Kategorien und Abdeckungsentscheidungen übernehmen. Vor einem produktiven Einsatz oder vor Anwendungen mit Auswirkungen auf reale Personen sollte der Inhalt daher sorgfältig geprüft werden.

Das Dataset sollte nicht als statistisch vollständige oder autoritative Darstellung Deutschlands, deutscher Regionen oder einzelner Bevölkerungsgruppen interpretiert werden.

Citation

If you find this dataset useful, you can cite it as:

@dataset{persona_german_2026,
  title = {Persona-German},
  author = {Ill-Ness and Selen Martin},
  year = {2026},
  month = {April},
  note = {Synthetic German persona dataset inspired by Nemotron-Personas-Korea},
  url = {https://huggingface.co/datasets/8F-ai/Persona-German}
}